Veles brings fast local code search to AI agents and developers

Veles is a local code search tool for AI agents and developers that combines BM25 keyword search with semantic search, and it does so entirely on CPU, according to its GitHub project page. The project is written in pure Rust and is designed to return results in tens of milliseconds from a persistent on-disk index.

The project is called Veles, and its author says it was originally inspired by Semble. Veles started as a Rust port of the same hybrid retrieval approach and has since added persistent and incremental indexing, tree-sitter-based symbol lookups, multiple output formats, glob and language filters, gRPC support, shell completions, and an MCP server for AI tools.

The core search mode blends BM25 and semantic scores using reciprocal rank fusion, or RRF, a ranking method that combines different result lists into one. The tool also uses an identifier-aware tokenizer that splits camelCase and snake_case names, and it can detect whether a query is likely asking for an exact symbol or a natural-language concept.

For code navigation, Veles supports tree-sitter commands for symbols, definitions, and references in Rust, Python, JavaScript, TypeScript, and Go. It also adds definition boosting, path penalties that push down test or compatibility files, and file saturation so one file does not dominate the results list.

The project includes a terminal UI called veles tui, which loads the cache once and updates searches as you type, with results in about 10 milliseconds per keystroke, according to the project description. The UI includes shortcuts to switch between hybrid, BM25, and semantic search, and it can show related code, definitions, and references for the current identifier.

Veles also ships with built-in MCP support over stdio, so tools like Claude and Cursor can call it through JSON-RPC. For service integrations, it exposes a gRPC server built on tonic with RPCs such as Index, Search, FindRelated, and GetStats.

The CLI can index a repo, refresh the cache after edits, and search with filters for language, path, exclude patterns, and minimum score. Output formats include pretty, compact, ripgrep-style, paths, JSON, and JSONL, which makes the tool usable both interactively and in scripts.

The project page says Veles keeps its index under <repo>/.veles/ and reuses embeddings for files that have not changed, based on size and modification time. It also lists prebuilt binaries for macOS, Linux, and Windows, plus install options through Homebrew, crates.io, and direct downloads from GitHub Releases.

According to the repository, the first search downloads a roughly 64 MB embedding model from Hugging Face and caches it locally. The project is released under the MIT license.